Estimating the growth rate of COVID-19

Using our analysis code

The main codebase to run our analysis is in analyze.py. The ExponentialGrowthRateEstimator takes in hyperparameters cumulative (Boolean) indicating whether the counts that will be passed in are cumulative counts or new cases. approximate_beta is used to interpolate missing data (days where the cumulative count is unknown). family is the generalized linear model family to use: either Poisson or NegativeBinomial. We recommend using NegativeBinomial, as the variance estimates from the Poisson seem to be a little optimistic, given how messy COVID-19 count data is. We recommend a default alpha value (for the Negative Binomial distribution) of 0.10 or 0.15.

Then, pass the numpy vectors day indicating the day of a measurement and cases indicating the respective count of new or cumulative cases, to fit.

The growth_rate and growth_rate_confint methods return the growth rate and 95% confidence intervals, respectively.

Example usage

A simple example of calling the analysis code is available in Simple Example.ipynb. See the Jupyter notebooks in Process States.ipynb or Main figures.ipynb for more complex examples of how we used this analysis to analyze and plot results for many regions at once.

Data sourcing

jhu_process_data.py will read the JHU data into: dict['country']['city']=array. The JHU repository is currently being reformatted so there may be some inconsistencies.

To pull the CSV files for global death data and state-level data, run the following two scripts:

sh pull_global.sh sh pull_states.sh

You'l probably want to run these roughly daily to keep the sources up to date.